PulseExploreJournal ClubDebatesTrendingResearchersJournals
Instagram
HomeExploreJournal ClubTrending
Synapse
⌘+K
Synapse
May 9, 2026Computers and Electronics in Agriculture0 citationsOpen Access

Centralized and vertical federated learning for Climate-Based crop yield prediction

View Full Paper
DKDaniyal Durmuş KöksalKBKwabena Ebo BenninBTBedir Tekinerdoğan

Key Points

  • This study evaluates and compares the effectiveness of Centralized Federated Learning (CFL) and Vertical Federated Learning (VFL) in predicting crop yield based on climate data.
  • Systematic comparison of CFL and VFL for modeling crop yields in Ethiopia.
  • Integration of ERA5 climate variables with maize and wheat yield simulations from LPJmL and DSSAT models.
  • Data remained at separate institutions in VFL, only exchanging latent representations.
  • CFL achieved predictive accuracy with R 2 ranging from 0.88 to 0.91; VFL yields R 2 from 0.77 to 0.90.
  • Both methods captured long-term yield trends and interannual variability as validated against FAO statistics.
  • VFL maintained stable convergence while exhibiting a modest performance reduction compared to CFL.

Abstract

• First systematic comparison of CFL and VFL for crop yield modeling. • CFL achieves R 2 = 0.88–0.91; VFL yields R 2 = 0.77–0.90 across crops. • VFL enables privacy-preserving prediction without raw data sharing. • ERA5 climate and LPJmL/DSSAT simulations integrated in federated setup. • FAO-validated over 15 years; both FL approaches capture yield trends. Accurate crop yield assessment is critical for food security planning and climate impact analysis, yet centralized machine learning approaches are increasingly constrained by data privacy requirements and institutional fragmentation. Climate variables, crop model simulations, and yield data are often distributed across organizations, making centralized data integration impractical. This study evaluates federated learning as a privacy-preserving alternative for large-scale crop yield modeling by systematically comparing Centralized Federated Learning (CFL) and Vertical Federated Learning (VFL). Using Ethiopia as a case study, we integrate ERA5 climate variables with maize and wheat yield simulations from the LPJmL and DSSAT models under irrigated and rainfed conditions. While CFL assumes centralized access to all features and labels, VFL employs a split neural network in which data remain at separate institutions and only latent representations are exchanged. Results show that CFL achieves the highest predictive accuracy (R 2 = 0.88–0.91), while VFL exhibits a modest performance reduction (R 2 = 0.765–0.897) but maintains stable convergence and minimal train–test divergence. Validation against FAO yield statistics over a 15-year period confirms that both approaches capture interannual variability and long-term trends. This work provides the first systematic empirical comparison of CFL and VFL for climate-driven crop yield modeling and demonstrates that vertical federated learning can deliver reliable predictions under realistic data-governance constraints, thereby enabling privacy-preserving agricultural yield assessment at scale.

Ask AI
Helpful
Bookmark
Share
View Full Paper

Cite This Study

Köksal et al. (2026) studied this question.

synapsesocial.com/papers/69fed19ab9154b0b82878edfhttps://doi.org/10.1016/j.compag.2026.111851
Ask AI
Helpful
Bookmark
Share
View Full Paper